kernel, linker, libc: Initial, probably horribly broken, support for some TLS situations

Makefile

@@ -101,6 +101,7 @@ KERNEL_OBJS += $(patsubst %.c,%.o,$(wildcard kernel/*/*/*.c))
 ##
 # Kernel objects from kernel/ assembly sources
 KERNEL_ASMOBJS = $(filter-out kernel/symbols.o,$(patsubst %.S,%.o,$(wildcard kernel/*.S)))
+HEADERS = $(wildcard base/usr/include/kernel/*.h base/usr/include/kernel/*/*.h)
 
 # Kernel
 
@@ -141,7 +142,6 @@ fatbase/mod:
 ##
 # Modules need to be installed on the boot image
 MODULES = $(patsubst modules/%.c,fatbase/mod/%.ko,$(wildcard modules/*.c))
-HEADERS = $(wildcard base/usr/include/kernel/*.h base/usr/include/kernel/*/*.h)
 
 fatbase/mod/%.ko: modules/%.c ${HEADERS} | fatbase/mod
 	${KCC} -nostdlib ${KCFLAGS} -c -o $@ $<

apps/test-tls.c

@@ -0,0 +1,37 @@
+#include <stdio.h>
+#include <pthread.h>
+#include <sys/sysfunc.h>
+
+__thread int myvalue;
+
+void * getaddressinthread(void * _unused) {
+	fprintf(stderr, "in thread before:\n");
+	fprintf(stderr, "&myvalue = %p\n", (void*)&myvalue);
+	fprintf(stderr, "myvalue  = %d\n", myvalue);
+	myvalue = 1234;
+	fprintf(stderr, "in thread after:\n");
+	fprintf(stderr, "&myvalue = %p\n", (void*)&myvalue);
+	fprintf(stderr, "myvalue  = %d\n", myvalue);
+	return NULL;
+}
+
+int main(int argc, char * argv[]) {
+
+	myvalue = 42;
+
+	fprintf(stderr, "main thread before:\n");
+	fprintf(stderr, "&myvalue = %p\n", (void*)&myvalue);
+	fprintf(stderr, "myvalue  = %d\n", myvalue);
+
+	pthread_t mythread;
+	pthread_create(&mythread, NULL, getaddressinthread, NULL);
+
+	void * retval;
+	pthread_join(mythread, &retval);
+
+	fprintf(stderr, "main thread after:\n");
+	fprintf(stderr, "&myvalue = %p\n", (void*)&myvalue);
+	fprintf(stderr, "myvalue  = %d\n", myvalue);
+
+	return 0;
+}

base/usr/include/kernel/process.h

@@ -35,6 +35,8 @@ typedef struct thread {
 	uint8_t    padding[32]; /* I don't know */
 
 	page_directory_t * page_directory; /* Page Directory */
+	uintptr_t  gsbase;
+
 } thread_t;
 
 /* Portable image struct */

base/usr/include/kernel/system.h

@@ -75,6 +75,8 @@ extern uint8_t startswith(const char * str, const char * accept);
 extern void gdt_install(void);
 extern void gdt_set_gate(uint8_t num, uint64_t base, uint64_t limit, uint8_t access, uint8_t gran);
 extern void set_kernel_stack(uintptr_t stack);
+extern void gdt_set_gsbase(uintptr_t base);
+extern uintptr_t gdt_get_gsbase(void);
 
 /* IDT */
 extern void idt_install(void);
@@ -86,7 +88,7 @@ extern void idt_set_gate(uint8_t num, void (*base)(void), uint16_t sel, uint8_t
  * the correct offsets as well.
  */
 struct regs {
-	unsigned int gs, fs, es, ds;
+	unsigned int _unused, fs, es, ds;
 	unsigned int edi, esi, ebp, esp, ebx, edx, ecx, eax;
 	unsigned int int_no, err_code;
 	unsigned int eip, cs, eflags, useresp, ss;

base/usr/include/pthread.h

@@ -12,6 +12,12 @@ typedef struct {
 } pthread_t;
 typedef unsigned int pthread_attr_t;
 
+typedef struct {
+	int volatile atomic_lock;
+	int volatile readers;
+	int writerPid;
+} pthread_rwlock_t;
+
 extern int pthread_create(pthread_t * thread, pthread_attr_t * attr, void *(*start_routine)(void *), void * arg);
 extern void pthread_exit(void * value);
 extern int pthread_kill(pthread_t thread, int sig);
@@ -38,5 +44,10 @@ extern int pthread_mutex_destroy(pthread_mutex_t *mutex);
 extern int pthread_attr_init(pthread_attr_t *attr);
 extern int pthread_attr_destroy(pthread_attr_t *attr);
 
+extern int pthread_rwlock_init(pthread_rwlock_t * lock, void * args);
+extern int pthread_rwlock_wrlock(pthread_rwlock_t * lock);
+extern int pthread_rwlock_rdlock(pthread_rwlock_t * lock);
+extern int pthread_rwlock_unlock(pthread_rwlock_t * lock);
+extern int pthread_rwlock_destroy(pthread_rwlock_t * lock);
 
 _End_C_Header

base/usr/include/sys/sysfunc.h

@@ -23,6 +23,7 @@
 #define TOARU_SYS_FUNC_THREADNAME   11
 #define TOARU_SYS_FUNC_DEBUGPRINT   12
 #define TOARU_SYS_FUNC_SETVGACURSOR 13
+#define TOARU_SYS_FUNC_SETGSBASE    14
 
 _Begin_C_Header
 extern int sysfunc(int command, char ** args);

kernel/cpu/gdt.c

@@ -30,9 +30,9 @@ typedef struct {
 
 /* In the future we may need to put a lock on the access of this */
 static struct {
-    gdt_entry_t entries[6];
+	gdt_entry_t entries[7];
-    gdt_pointer_t pointer;
+	gdt_pointer_t pointer;
-    tss_entry_t tss;
+	tss_entry_t tss;
 } gdt __attribute__((used));
 
 extern void gdt_flush(uintptr_t);
@@ -53,6 +53,17 @@ void gdt_set_gate(uint8_t num, uint64_t base, uint64_t limit, uint8_t access, ui
 	ENTRY(num).access = access;
 }
 
+void gdt_set_gsbase(uintptr_t base) {
+	ENTRY(6).base_low = (base & 0xFFFF);
+	ENTRY(6).base_middle = (base >> 16) & 0xFF;
+	ENTRY(6).base_high = (base >> 24) & 0xFF;
+	asm volatile ("mov %0, %%gs" :: "r"((6 << 3) | 0x3));
+}
+
+uintptr_t gdt_get_gsbase(void) {
+	return (ENTRY(6).base_low) | (ENTRY(6).base_middle << 16) | (ENTRY(6).base_high << 24);
+}
+
 static void write_tss(int32_t num, uint16_t ss0, uint32_t esp0);
 
 void gdt_install(void) {
@@ -65,8 +76,8 @@ void gdt_install(void) {
 	gdt_set_gate(2, 0, 0xFFFFFFFF, 0x92, 0xCF); /* Data segment */
 	gdt_set_gate(3, 0, 0xFFFFFFFF, 0xFA, 0xCF); /* User code */
 	gdt_set_gate(4, 0, 0xFFFFFFFF, 0xF2, 0xCF); /* User data */
-
 	write_tss(5, 0x10, 0x0);
+	gdt_set_gate(6, 0, 0xFFFFFFFF, 0xF2, 0xCF);
 
 	/* Go go go */
 	gdt_flush((uintptr_t)gdtp);

kernel/sys/syscall.c

@@ -784,6 +784,12 @@ static int sys_sysfunc(int fn, char ** args) {
 
 			return 0;
 
+		case 14:
+			PTR_VALIDATE(args);
+			current_process->thread.gsbase = (uintptr_t)args[0];
+			gdt_set_gsbase(current_process->thread.gsbase);
+			return 0;
+
 		default:
 			debug_print(ERROR, "Bad system function %d", fn);
 			break;

kernel/sys/task.c

@@ -218,6 +218,7 @@ uint32_t fork(void) {
 
 	new_proc->thread.esp = esp;
 	new_proc->thread.ebp = ebp;
+	new_proc->thread.gsbase = current_process->thread.gsbase;
 
 	new_proc->is_tasklet = parent->is_tasklet;
 
@@ -272,6 +273,7 @@ int create_kernel_tasklet(tasklet_t tasklet, char * name, void * argp) {
 
 	new_proc->thread.esp = esp;
 	new_proc->thread.ebp = ebp;
+	new_proc->thread.gsbase = current_process->thread.gsbase;
 
 	new_proc->thread.eip = (uintptr_t)tasklet;
 
@@ -339,6 +341,7 @@ clone(uintptr_t new_stack, uintptr_t thread_func, uintptr_t arg) {
 
 	new_proc->thread.esp = esp;
 	new_proc->thread.ebp = ebp;
+	new_proc->thread.gsbase = current_process->thread.gsbase;
 
 	new_proc->is_tasklet = parent->is_tasklet;
 
@@ -405,6 +408,7 @@ void switch_task(uint8_t reschedule) {
 	current_process->thread.eip = eip;
 	current_process->thread.esp = esp;
 	current_process->thread.ebp = ebp;
+	current_process->thread.gsbase = gdt_get_gsbase();
 	current_process->running = 0;
 
 	/* Save floating point state */
@@ -432,6 +436,7 @@ void switch_next(void) {
 	eip = current_process->thread.eip;
 	esp = current_process->thread.esp;
 	ebp = current_process->thread.ebp;
+	gdt_set_gsbase(current_process->thread.gsbase);
 	unswitch_fpu();
 
 	/* Validate */

kernel/user.S

@@ -32,6 +32,8 @@ enter_userspace:
     mov %eax, %ds
     mov %eax, %es
     mov %eax, %fs
+
+    mov $0x33, %ax
     mov %eax, %gs
     /* %ss is handled by iret */
 

libc/main.c

@@ -30,8 +30,11 @@ void _exit(int val){
 	__builtin_unreachable();
 }
 
+extern void __make_tls(void);
+
 __attribute__((constructor))
 static void _libc_init(void) {
+	__make_tls();
 	__stdio_init_buffers();
 
 	unsigned int x = 0;

libc/pthread/pthread.c

@@ -11,6 +11,7 @@
 #include <errno.h>
 
 #include <sys/wait.h>
+#include <sys/sysfunc.h>
 
 DEFN_SYSCALL3(clone, SYS_CLONE, uintptr_t, uintptr_t, void *);
 DEFN_SYSCALL0(gettid, SYS_GETTID);
@@ -24,11 +25,39 @@ int gettid() {
 	return syscall_gettid(); /* never fails */
 }
 
+struct pthread {
+	void * (*entry)(void *);
+	void * arg;
+};
+
+void * ___tls_get_addr(void* input) {
+	return NULL;
+}
+
+void __make_tls(void) {
+	char * tlsSpace = calloc(1,4096);
+	char ** tlsSelf  = (char **)(tlsSpace + 4096 - sizeof(char *));
+	*tlsSelf = (char*)tlsSelf;
+	sysfunc(TOARU_SYS_FUNC_SETGSBASE, (char*[]){(char*)tlsSelf});
+}
+
+void * __thread_start(void * thread) {
+	__make_tls();
+	struct pthread * me = ((pthread_t *)thread)->ret_val;
+	((pthread_t *)thread)->ret_val = 0;
+	return me->entry(me->arg);
+}
+
 int pthread_create(pthread_t * thread, pthread_attr_t * attr, void *(*start_routine)(void *), void * arg) {
 	char * stack = malloc(PTHREAD_STACK_SIZE);
 	uintptr_t stack_top = (uintptr_t)stack + PTHREAD_STACK_SIZE;
+
 	thread->stack = stack;
-	thread->id = clone(stack_top, (uintptr_t)start_routine, arg);
+	struct pthread * data = malloc(sizeof(struct pthread));
+	data->entry = start_routine;
+	data->arg   = arg;
+	thread->ret_val = data;
+	thread->id = clone(stack_top, (uintptr_t)__thread_start, thread);
 	return 0;
 }
 

libc/pthread/pthread_rwlock.c

@@ -0,0 +1,61 @@
+#include <stdlib.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <syscall.h>
+#include <syscall_nums.h>
+#include <signal.h>
+#include <pthread.h>
+#include <errno.h>
+
+#include <sys/wait.h>
+
+#define ACQUIRE_LOCK() do { while (__sync_lock_test_and_set(&lock->atomic_lock, 0x01)) { syscall_yield(); } } while (0)
+#define RELEASE_LOCK() do { __sync_lock_release(&lock->atomic_lock); } while (0)
+
+int pthread_rwlock_init(pthread_rwlock_t * lock, void * args) {
+	lock->readers = 0;
+	lock->atomic_lock = 0;
+	if (args != NULL) {
+		fprintf(stderr, "pthread: pthread_rwlock_init arg unsupported\n");
+		return 1;
+	}
+	return 0;
+}
+
+int pthread_rwlock_wrlock(pthread_rwlock_t * lock) {
+	ACQUIRE_LOCK();
+	while (1) {
+		if (lock->readers == 0) {
+			lock->readers = -1;
+			lock->writerPid = syscall_getpid();
+			RELEASE_LOCK();
+			return 0;
+		}
+		syscall_yield();
+	}
+}
+
+int pthread_rwlock_rdlock(pthread_rwlock_t * lock) {
+	ACQUIRE_LOCK();
+	while (1) {
+		if (lock->readers >= 0) {
+			lock->readers++;
+			RELEASE_LOCK();
+			return 0;
+		}
+		syscall_yield();
+	}
+}
+
+int pthread_rwlock_unlock(pthread_rwlock_t * lock) {
+	ACQUIRE_LOCK();
+	if (lock->readers > 0) lock->readers--;
+	else if (lock->readers < 0) lock->readers = 0;
+	else fprintf(stderr, "pthread: bad lock state detected\n");
+	RELEASE_LOCK();
+	return 0;
+}
+
+int pthread_rwlock_destroy(pthread_rwlock_t * lock) {
+	return 0;
+}

libc/stdlib/realpath.c

@@ -41,9 +41,7 @@ char *realpath(const char *path, char *resolved_path) {
 	}
 
 	if (!resolved_path) {
-		/* Can't support this yet. */
+		resolved_path = malloc(PATH_MAX+1);
-		errno = -EINVAL;
-		return NULL;
 	}
 
 	/* If we're lucky, we can do this with no allocations, so let's start here... */

linker/linker.c

@@ -75,6 +75,8 @@ typedef int (*entry_point_t)(int, char *[], char**);
 static hashmap_t * dumb_symbol_table;
 static hashmap_t * glob_dat;
 static hashmap_t * objects_map;
+static hashmap_t * tls_map;
+static size_t current_tls_offset = 0;
 
 /* Used for dlerror */
 static char * last_error = NULL;
@@ -388,6 +390,7 @@ static int need_symbol_for_type(unsigned char type) {
 		case 5:
 		case 6:
 		case 7:
+		case 14:
 			return 1;
 		default:
 			return 0;
@@ -471,6 +474,20 @@ static int object_relocate(elf_t * object) {
 					case 5: /* COPY */
 						memcpy((void *)(table->r_offset + object->base), (void *)x, sym->st_size);
 						break;
+					case 14: /* TLS_TPOFF */
+						x = *((ssize_t *)(table->r_offset + object->base));
+						if (!hashmap_has(tls_map, symname)) {
+							if (!sym->st_size) {
+								fprintf(stderr, "Haven't placed %s in static TLS yet but don't know its size?\n", symname);
+							}
+							current_tls_offset += sym->st_size; /* TODO alignment restrictions */
+							hashmap_set(tls_map, symname, (void*)(current_tls_offset));
+							x -= current_tls_offset;
+						} else {
+							x -= (size_t)hashmap_get(tls_map, symname);
+						}
+						memcpy((void *)(table->r_offset + object->base), &x, sizeof(uintptr_t));
+						break;
 					default:
 						TRACE_LD("Unknown relocation type: %d", type);
 				}
@@ -751,6 +768,7 @@ int main(int argc, char * argv[]) {
 	dumb_symbol_table = hashmap_create(10);
 	glob_dat = hashmap_create(10);
 	objects_map = hashmap_create(10);
+	tls_map = hashmap_create(10);
 
 	/* Setup symbols for built-in exports */
 	ld_exports_t * ex = ld_builtin_exports;